Precise forecasting of photovoltaic (PV) power is crucial for maintaining grid stability and effectively incorporating solar energy. This study compares four forecasting models, LSTM, GRU, Random Forest, and XGBoost, using three years of real operational data (2023–2025) from the DKASC solar database. Fourteen input features were constructed from irradiance measurements, meteorological variables, rainfall, wind direction, and cyclical time features. A sliding window of 36 time steps captured temporal dependencies, while the dataset was chronologically split into training, validation, and testing subsets. Model performance was assessed using R², RMSE, MAE, and MAPE. Results show LSTM achieved the highest accuracy with R² of 0.9647 and RMSE of 88.40 W, whereas XGBoost obtained the lowest MAPE of 19.18%. SHAP analysis and an ablation study were conducted to identify influential features and evaluate their contributions to forecasting performance. SHAP analysis identified global horizontal radiation, global tilted radiation, and cyclical temporal features, particularly day_cos and hour_cos, as the most influential predictors, while wind speed and daily rainfall showed limited impact. The findings demonstrate that deep learning models can provide reliable PV power forecasts when combined with informative feature engineering and appropriate temporal window selection.
Copyrights © 2026